U.S. County-Level Utilization Forecasting
- The model runs a modified version of Penn Medicine's CHIME (COVID-19 Hospital Impact Model for Epidemics) Model
- It normally uses a hospital market share parameter.
- We treat each county as a single hospital with a market share of 100%.
- Case counts in each county are modeled separately (3,144 models run).
- Each county is treated as an independent population, unless the current case count is less than 10. If so, neighboring counties (chosen via proximity) are added to the model population until the case count is at least 10. The model outputs for peak predictions are then pro-rated for county population.
- Parameters are tuned using values from literature review.
- Public Data Sources are updated at 7 p.m. CT daily (reporting for the day prior). The Cerner model is then updated the following morning at 8 a.m. CT.
- County level and upcoming peaks data file downloads:
- Results are not guaranteed.
Only publicly available data sources were used to build the Cerner model. The public data sources include:
- Definitive Healthcare hospital bed data
- Johns Hopkins CSSE COVID-19 case count data
- US Census county population counts
- Dataset from American Red Cross: COVID19 Public Health Emergency Status by County
- hospitalized_rate=0.2, icu_rel_rate=0.21, vent_rel_rate=0.9
- hospitalized_rate: proportion of confirmed COVID-19 positive patients hospitalized
- icu_rel_rate: proportion of hospitalized COVID-19 patients transferred to ICU
- vent_rel_rate: proportion of COVID-19 ICU patients who need ventilation
- Based on empirical observations at Cerner Clients and literature review
- hospital_days=7, icu_days=9, ventilated_days=10
- Number of days spent in hospital/ICU/ventilator. Based on literature review, conservative estimates used
- Based on literature review, estimates range from 15% to 80%
- County population based on US Census data
- Based on literature review
- Set to 100% since we treat each county as a single hospital
- Estimation of effectiveness of social distancing (higher is more reduction). Start date of social distancing from Red Cross dataset.
- Estimated days for case count to double, estimated based on time series data of epidemic in the first 20 days when the case count was at least 5 to avoid picking up the effects of social distancing and outlier effects when the case count is very low.
- Error on peak dates is about +/- 1-2 weeks, unless R_0 is near 1. When R_0 is near 1, especially when due to social distancing, it becomes very unclear whether the peak is in the past (for example, R_0=0.99), or far in the future (for example, R_0=1.01). Because of the model implementation, peaks predicted to be 6+ weeks in future often result from this situation.
- Peak utilization ratios are correct to about a factor of 2 (between twice and half the real number).
- Model is most sensitive to doubling_time and relative_contact_ratio.
- Estimates for most counties are later than state-wide models, which are generally most representative of urban areas.